Spectral Tensor Synthesis Analysis for Hyperspectral Image Spectral–Spatial Feature Extraction

Feature extraction is a preprocessing step for hyperspectral image classification. Principal component analysis only uses the spectral information, but it does not use spatial information of a hyperspectral image. Both spatial and spectral information are used when hyperspectral image is modelled as tensor, that is, decreasing the noise on spatial dimension and reducing the dimension on a spectral dimension at the same time. However, in this model, a hyperspectral image is modelled only as a data cube. The factors affecting the spectral features of ground objects is not considered and these factors are barely distinguished. This means that further improving classification is very difficult. Therefore, a new model on hyperspectral image is proposed by the authors. In the new model, many factors that impact the spectral features of ground objects are synthesized as the within-class factor. The within-class factor, the class factor and the pixel spectral are selected as a mode, respectively. The pixel spectrals in the training set are modelled as a third-order tensor. The experiment results indicate that the new method improves the classification compared with the previous methods.